Compact Web Video Summarization Via Supervised Learning

Ever growing consumption of online videos from search, recommendation, and sharing has generated a strong demand on compact summarization, to allow users to quickly understand the video content and make the whether-to-watch decision. This paper explores achieving this using a compact set of four thumbnails, via supervised learning methods. Due to the ubiquitous disagreements among the thumbnail sets preferred by different labelers, there exists no unique ground truth set. To address this problem, we propose the pair wise ranking method, which trains the model to best predict the user preference over each pair of thumbnail set candidates. Experimental results on a large video dataset showed that the proposed method outperforms the existing schemes by a large margin.